model.py

# --------------------------------------------------------
# References:
# timm: https://github.com/rwightman/pytorch-image-models/tree/master/timm
# DeiT: https://github.com/facebookresearch/deit
# --------------------------------------------------------
 
from functools import partial

import torch
from torch._C import Value
import torch.nn as nn
import numpy as np

from timm.models.vision_transformer import PatchEmbed, Block
from transformers import EncoderDecoderModel, BertTokenizer, AutoTokenizer


from torch import einsum, nn
import torch.nn.functional as F
from einops import rearrange, repeat

import torch
import torch.nn as nn
import torch.nn.functional as F

class FocalLoss(nn.CrossEntropyLoss):
    ''' Focal loss for classification tasks on imbalanced datasets '''

    def __init__(self, gamma=1.0, alpha=None, ignore_index=-100, reduction='none'):
        super().__init__(weight=alpha, ignore_index=ignore_index, reduction='none')
        self.reduction = reduction
        self.gamma = gamma

    def forward(self, input_, target):
        cross_entropy = super().forward(input_, target)
        # Temporarily mask out ignore index to '0' for valid gather-indices input.
        # This won't contribute final loss as the cross_entropy contribution
        # for these would be zero.
        target = target * (target != self.ignore_index).long()
        input_prob = torch.gather(F.softmax(input_, 1), 1, target.unsqueeze(1)).squeeze(1)
        loss = torch.pow(1 - input_prob, self.gamma) * cross_entropy
        return torch.mean(loss) if self.reduction == 'mean' \
               else torch.sum(loss) if self.reduction == 'sum' \
               else loss


# helper functions

import math
from functools import reduce

def prob_mask_like(t, prob):
    return torch.zeros_like(t).float().uniform_(0, 1) < prob

def mask_with_tokens(t, token_ids):
    init_no_mask = torch.full_like(t, False, dtype=torch.bool)
    mask = reduce(lambda acc, el: acc | (t == el), token_ids, init_no_mask)
    return mask

def get_mask_subset_with_prob(mask, prob):
    batch, seq_len, device = *mask.shape, mask.device
    max_masked = math.ceil(prob * seq_len)

    num_tokens = mask.sum(dim=-1, keepdim=True)
    mask_excess = (mask.cumsum(dim=-1) > (num_tokens * prob).ceil())
    mask_excess = mask_excess[:, :max_masked]

    rand = torch.rand((batch, seq_len), device=device).masked_fill(~mask, -1e9)
    _, sampled_indices = rand.topk(max_masked, dim=-1)
    sampled_indices = (sampled_indices + 1).masked_fill_(mask_excess, 0)

    new_mask = torch.zeros((batch, seq_len + 1), device=device)
    new_mask.scatter_(-1, sampled_indices, 1)
    return new_mask[:, 1:].bool()


def exists(val):
    return val is not None

def default(val, d):
    return val if exists(val) else d

# normalization
# they use layernorm without bias, something that pytorch does not offer


class LayerNorm(nn.Module):
    def __init__(self, dim):
        super().__init__()
        self.gamma = nn.Parameter(torch.ones(dim))
        self.register_buffer("beta", torch.zeros(dim))

    def forward(self, x):
        return F.layer_norm(x, x.shape[-1:], self.gamma, self.beta)

# residual
class Residual(nn.Module):
    def __init__(self, fn):
        super().__init__()
        self.fn = fn

    def forward(self, x, *args, **kwargs):
        return self.fn(x, *args, **kwargs) + x

# rotary positional embedding
# https://arxiv.org/abs/2104.09864
class RotaryEmbedding(nn.Module):
    def __init__(self, dim):
        super().__init__()
        inv_freq = 1.0 / (10000 ** (torch.arange(0, dim, 2).float() / dim))
        self.register_buffer("inv_freq", inv_freq)

    def forward(self, max_seq_len, *, device):
        seq = torch.arange(max_seq_len, device=device, dtype=self.inv_freq.dtype)
        freqs = einsum("i , j -> i j", seq, self.inv_freq)
        return torch.cat((freqs, freqs), dim=-1)


def rotate_half(x):
    x = rearrange(x, "... (j d) -> ... j d", j=2)
    x1, x2 = x.unbind(dim=-2)
    return torch.cat((-x2, x1), dim=-1)


def apply_rotary_pos_emb(pos, t):
    return (t * pos.cos()) + (rotate_half(t) * pos.sin())


# classic Noam Shazeer paper, except here they use SwiGLU instead of the more popular GELU for gating the feedforward
# https://arxiv.org/abs/2002.05202
class SwiGLU(nn.Module):
    def forward(self, x):
        x, gate = x.chunk(2, dim=-1)
        return F.silu(gate) * x


# parallel attention and feedforward with residual
# discovered by Wang et al + EleutherAI from GPT-J fame
class ParallelTransformerBlock(nn.Module):
    def __init__(self, dim, dim_head=64, heads=8, ff_mult=4, attn_drop_rate=0.0):
        super().__init__()
        self.norm = LayerNorm(dim)

        attn_inner_dim = dim_head * heads
        ff_inner_dim = dim * ff_mult
        self.fused_dims = (attn_inner_dim, dim_head, dim_head, (ff_inner_dim * 2))

        self.heads = heads
        self.scale = dim_head**-0.5
        self.rotary_emb = RotaryEmbedding(dim_head)

        self.fused_attn_ff_proj = nn.Linear(dim, sum(self.fused_dims), bias=False)
        self.attn_out = nn.Linear(attn_inner_dim, dim, bias=False)

        self.ff_out = nn.Sequential(
            SwiGLU(),
            nn.Linear(ff_inner_dim, dim, bias=False)
        )

        self.attn_drop_rate = attn_drop_rate

        # for caching causal mask and rotary embeddings

        self.register_buffer("mask", None, persistent=False)
        self.register_buffer("pos_emb", None, persistent=False)

    def get_mask(self, n, device):
        if self.mask is not None and self.mask.shape[-1] >= n:
            return self.mask[:n, :n]

        mask = torch.ones((n, n), device=device, dtype=torch.bool).triu(1)
        self.register_buffer("mask", mask, persistent=False)
        return mask

    def get_rotary_embedding(self, n, device):
        if self.pos_emb is not None and self.pos_emb.shape[-2] >= n:
            return self.pos_emb[:n]

        pos_emb = self.rotary_emb(n, device=device)
        self.register_buffer("pos_emb", pos_emb, persistent=False)
        return pos_emb

    def forward(self, x, attn_mask=None):
        """
        Performs self attention and feedforward
        einstein notation
        b - batch
        h - heads
        n, i, j - sequence length (base sequence length, source, target)
        d - feature dimension
        """

        n, device, h = x.shape[1], x.device, self.heads
        # pre layernorm
        x = self.norm(x)
        # attention queries, keys, values, and feedforward inner
        q, k, v, ff = self.fused_attn_ff_proj(x).split(self.fused_dims, dim=-1)

        # split heads
        # they use multi-query single-key-value attention, yet another Noam Shazeer paper
        # they found no performance loss past a certain scale, and more efficient decoding obviously
        # https://arxiv.org/abs/1911.02150
        q = rearrange(q, "b n (h d) -> b h n d", h=h)
        # rotary embeddings
        positions = self.get_rotary_embedding(n, device)
        q, k = map(lambda t: apply_rotary_pos_emb(positions, t), (q, k))
        # scale
        q = q * self.scale
        # similarity
        sim = einsum("b h i d, b j d -> b h i j", q, k)
        # causal mask
        causal_mask = self.get_mask(n, device)
        sim = sim.masked_fill(causal_mask, -torch.finfo(sim.dtype).max)

        # extra attention mask - for masking out attention from text CLS token to padding
        if exists(attn_mask):
            attn_mask = rearrange(attn_mask, 'b i j -> b 1 i j')
            sim = sim.masked_fill(~attn_mask, -torch.finfo(sim.dtype).max)

            if self.attn_drop_rate != 0.:
                # import ipdb; ipdb.set_trace()
                drop_ind = sim != -torch.finfo(sim.dtype).max
                dropout_mask = torch.cuda.FloatTensor(*sim[drop_ind].shape).uniform_() > self.attn_drop_rate
                sim[drop_ind] = sim[drop_ind].masked_fill(~dropout_mask, -torch.finfo(sim.dtype).max)

        # attention
        sim = sim - sim.amax(dim=-1, keepdim=True).detach()
        attn = sim.softmax(dim=-1)
        # aggregate values
        out = einsum("b h i j, b j d -> b h i d", attn, v)
        # merge heads
        out = rearrange(out, "b h n d -> b n (h d)")
        return self.attn_out(out) + self.ff_out(ff)

# cross attention - using multi-query + one-headed key / values as in PaLM w/ optional parallel feedforward
class CrossAttention(nn.Module):
    def __init__(
        self,
        dim,
        *,
        context_dim=None,
        dim_head=64,
        heads=8,
        parallel_ff=False,
        ff_mult=4,
        norm_context=False,
        dropout=0.0,
    ):
        super().__init__()
        self.heads = heads
        self.scale = dim_head ** -0.5
        inner_dim = heads * dim_head
        context_dim = default(context_dim, dim)

        self.norm = LayerNorm(dim)
        self.context_norm = LayerNorm(context_dim) if norm_context else nn.Identity()

        self.to_q = nn.Linear(dim, inner_dim, bias=False)
        self.to_kv = nn.Linear(context_dim, dim_head * 2, bias=False)
        self.to_out = nn.Linear(inner_dim, dim, bias=False)

        self.dropout = dropout

        # whether to have parallel feedforward
        ff_inner_dim = ff_mult * dim

        self.ff = nn.Sequential(
            nn.Linear(dim, ff_inner_dim * 2, bias=False),
            SwiGLU(),
            nn.Linear(ff_inner_dim, dim, bias=False)
        ) if parallel_ff else None

    def forward(self, x, context):
        """
        Use text and query, and image as kv
        einstein notation
        b - batch
        h - heads
        n, i, j - sequence length (base sequence length, source, target)
        d - feature dimension
        """

        # pre-layernorm, for queries and context
        x = self.norm(x)
        context = self.context_norm(context)
        # get queries
        q = self.to_q(x)
        q = rearrange(q, 'b n (h d) -> b h n d', h = self.heads)
        # scale
        q = q * self.scale
        # get key / values
        k, v = self.to_kv(context).chunk(2, dim=-1)
        # query / key similarity
        sim = einsum('b h i d, b j d -> b h i j', q, k)
        
        # dropout
        if self.training:
            dropout_mask = torch.cuda.FloatTensor(*sim.shape).uniform_() > self.dropout
            sim = sim.masked_fill(~dropout_mask, -torch.finfo(sim.dtype).max)

        # attention
        sim = sim - sim.amax(dim=-1, keepdim=True)
        attn = sim.softmax(dim=-1)
        # aggregate
        out = einsum('b h i j, b j d -> b h i d', attn, v)
        # merge and combine heads
        out = rearrange(out, 'b h n d -> b n (h d)')
        out = self.to_out(out)
        # add parallel feedforward (for multimodal layers)
        if exists(self.ff):
            out = out + self.ff(x)
        return out



def get_2d_sincos_pos_embed(embed_dim, grid_size, cls_token=False):
    """
    grid_size: int of the grid height and width
    return:
    pos_embed: [grid_size*grid_size, embed_dim] or [1+grid_size*grid_size, embed_dim] (w/ or w/o cls_token)
    """
    grid_h = np.arange(grid_size, dtype=np.float32)
    grid_w = np.arange(grid_size, dtype=np.float32)
    grid = np.meshgrid(grid_w, grid_h)  # here w goes first
    grid = np.stack(grid, axis=0)

    grid = grid.reshape([2, 1, grid_size, grid_size])
    pos_embed = get_2d_sincos_pos_embed_from_grid(embed_dim, grid)
    if cls_token:
        pos_embed = np.concatenate([np.zeros([1, embed_dim]), pos_embed], axis=0)
    return pos_embed

def get_2d_sincos_pos_embed_from_grid(embed_dim, grid):
    assert embed_dim % 2 == 0

    # use half of dimensions to encode grid_h
    emb_h = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[0])  # (H*W, D/2)
    emb_w = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[1])  # (H*W, D/2)

    emb = np.concatenate([emb_h, emb_w], axis=1) # (H*W, D)
    return emb

def get_1d_sincos_pos_embed_from_grid(embed_dim, pos):
    """
    embed_dim: output dimension for each position
    pos: a list of positions to be encoded: size (M,)
    out: (M, D)
    """
    assert embed_dim % 2 == 0
    omega = np.arange(embed_dim // 2, dtype=np.float32)
    omega /= embed_dim / 2.
    omega = 1. / 10000**omega  # (D/2,)

    pos = pos.reshape(-1)  # (M,)
    out = np.einsum('m,d->md', pos, omega)  # (M, D/2), outer product

    emb_sin = np.sin(out) # (M, D/2)
    emb_cos = np.cos(out) # (M, D/2)

    emb = np.concatenate([emb_sin, emb_cos], axis=1)  # (M, D)
    return emb

class MaskedAutoencoderViT(nn.Module):
    """ Masked Autoencoder with VisionTransformer backbone
    """
    def __init__(self, img_size=224, patch_size=16, in_chans=3,
                 embed_dim=1024, depth=24, num_heads=16,
                 decoder_embed_dim=512, decoder_depth=8, decoder_num_heads=16,
                 mlp_ratio=4., norm_layer=nn.LayerNorm, norm_pix_loss=True,
                 unimodal_depth=2, multimodal_depth=8, dim_head=64,heads=8,
                 ff_mult=4, extract_multi_level=False, use_focal_loss=False, focal_gamma=1.0,
                 less_u=False, use_weak_negative=False, use_label_smooth=False, ls_coef=0.1,
                 use_maximum_entropy=False, ce_additional=False, use_word_weights=False, use_token_pos=False,
                 use_expect_k=False, use_top_k=False, mae_decoder_caption=False, decoder_slot_depth=2, disable_decoder_vis_token_grad=False,
                 cross_attn_dropout=0.0, predict_next_k_words=False, next_k=3, masked_text=False, masked_text_ratio=0.25, text_length=70,
                 projector_layer=0, uni_dim=1024, uni_dim_head=64, uni_heads=8, uni_ff_mult=4, text_drop_attn=0.):
        super().__init__()

        # --------------------------------------------------------------------------
        # MAE encoder specifics
        self.patch_embed = PatchEmbed(img_size, patch_size, in_chans, embed_dim)
        num_patches = self.patch_embed.num_patches

        self.cls_token = nn.Parameter(torch.zeros(1, 1, embed_dim))
        self.pos_embed = nn.Parameter(torch.zeros(1, num_patches + 1, embed_dim), requires_grad=False)  # fixed sin-cos embedding

        self.blocks = nn.ModuleList([
            Block(embed_dim, num_heads, mlp_ratio, qkv_bias=True, norm_layer=norm_layer)
            for i in range(depth)])
        self.norm = norm_layer(embed_dim)
        # --------------------------------------------------------------------------

        # --------------------------------------------------------------------------
        # MAE decoder specifics
        self.decoder_embed = nn.Linear(embed_dim, decoder_embed_dim, bias=True)

        self.mask_token = nn.Parameter(torch.zeros(1, 1, decoder_embed_dim))

        self.decoder_pos_embed = nn.Parameter(torch.zeros(1, num_patches + 1, decoder_embed_dim), requires_grad=False)  # fixed sin-cos embedding

        self.mae_decoder_depth = decoder_depth
        self.mae_decoder_caption = mae_decoder_caption
        self.decoder_blocks = nn.ModuleList([
            Block(decoder_embed_dim, decoder_num_heads, mlp_ratio, qkv_bias=True, norm_layer=norm_layer)
            for i in range(decoder_depth)])

        if self.mae_decoder_caption:

            self.decoder_slot_layers = nn.ModuleList([])
            for _ in range(decoder_slot_depth):
                self.decoder_slot_layers.append(
                    Residual(CrossAttention(dim=decoder_embed_dim, dim_head=dim_head, heads=heads, parallel_ff=True, ff_mult=ff_mult,)),
                    # Residual(CrossAttention(dim=decoder_embed_dim, dim_head=dim_head, heads=heads, parallel_ff=True, ff_mult=ff_mult,))
                )
            self.decoder_caption_proj = nn.Linear(decoder_embed_dim, embed_dim)
            self.disable_decoder_vis_token_grad = disable_decoder_vis_token_grad

        self.decoder_norm = norm_layer(decoder_embed_dim)
        self.decoder_pred = nn.Linear(decoder_embed_dim, patch_size**2 * in_chans, bias=True) # encoder to decoder
        # --------------------------------------------------------------------------

        self.norm_pix_loss = norm_pix_loss

        # --------------------------------------------------------------------------
        # captioner specifics
        # unimodal layer is for text tokens.
        # multimodal layer is for text to query from image.        
        self.tokenizer = AutoTokenizer.from_pretrained("bert-base-uncased", )
        
        # token embeddings
        # NOTE: +1 for mask token used by MLM objective
        # self.token_emb = nn.Embedding(len(self.tokenizer.vocab) + 1, uni_dim)

        self.token_emb = nn.Embedding(len(self.tokenizer.vocab), uni_dim)
        self.text_cls_token = nn.Parameter(torch.randn(uni_dim))

        self.embed_dim = embed_dim
        self.uni_dim = uni_dim
        
        #import ipdb; ipdb.set_trace()
        # unimodal layers
        # TODO: search on the four parameters
        # uni_dim=1024, uni_dim_head=64, uni_heads=8, uni_ff_mult=4
        self.text_drop_attn = text_drop_attn
        self.unimodal_layers = nn.ModuleList([])
        for _ in range(unimodal_depth):
            self.unimodal_layers.append(
                Residual(ParallelTransformerBlock(dim=uni_dim, dim_head=uni_dim_head, 
                        heads=uni_heads, ff_mult=uni_ff_mult, attn_drop_rate=self.text_drop_attn)),
            )

        self.need_uni_2_mul_proj = False
        if uni_dim != embed_dim:
            self.need_uni_2_mul_proj = True
            self.uni_2_mul_proj = nn.Linear(uni_dim, embed_dim)
        
        # multimodal layers
        self.multimodal_layers = nn.ModuleList([])
        self.less_u = less_u
        if less_u:
            for _ in range(multimodal_depth):
                self.multimodal_layers.append(nn.ModuleList([
                    Residual(CrossAttention(dim=embed_dim, dim_head=dim_head, heads=heads, parallel_ff=True, ff_mult=ff_mult, dropout=cross_attn_dropout)),
                    Residual(CrossAttention(dim=embed_dim, dim_head=dim_head, heads=heads, parallel_ff=True, ff_mult=ff_mult, dropout=cross_attn_dropout))
                ]))
        else:
            for _ in range(multimodal_depth):
                self.multimodal_layers.append(nn.ModuleList([
                    Residual(ParallelTransformerBlock(dim=embed_dim, dim_head=dim_head, heads=heads, ff_mult=ff_mult)),
                    Residual(CrossAttention(dim=embed_dim, dim_head=dim_head, heads=heads, parallel_ff=True, ff_mult=ff_mult, dropout=cross_attn_dropout))
                ]))

        # to logits: for softmax caption loss
        self.to_logits = nn.Sequential(
            LayerNorm(embed_dim),
            nn.Linear(embed_dim, len(self.tokenizer.vocab), bias=False)
        )

        self.ce_additional = ce_additional
        if ce_additional:
            # to logits: for other losses
            self.to_logits_1 = nn.Sequential(
                LayerNorm(embed_dim),
                nn.Linear(embed_dim, len(self.tokenizer.vocab), bias=False)
            )
        
        nn.init.normal_(self.token_emb.weight, std=0.02)

        self.pad_id = 0
        self.cls_id = 101
        self.sep_id = 102

        self.logsoftmax = nn.LogSoftmax(dim=1)

        self.extract_multi_level = extract_multi_level
        if self.extract_multi_level:
            self.projectors = nn.ModuleList([nn.Sequential(
                nn.Linear(embed_dim, embed_dim // 2),
                nn.GELU(),
                nn.Linear(embed_dim // 2, embed_dim),
                norm_layer(embed_dim)
            ) for _ in [2, 5, 8,]])
        # --------------------------------------------------------------------------
        
        self.use_focal_loss = use_focal_loss
        
        self.use_weak_negative = use_weak_negative
        self.use_label_smooth = use_label_smooth
        self.ls_coef = ls_coef
        self.use_entropy = use_maximum_entropy
        self.use_word_weights = use_word_weights
        self.use_token_pos = use_token_pos

        self.predict_next_k_words = predict_next_k_words
        self.next_k = next_k
        self.pad = torch.nn.ReplicationPad1d((0, self.next_k-1))

        self.use_expect_k = use_expect_k
        self.use_top_k = use_top_k

        if self.use_word_weights or self.use_token_pos:
            self.focal_loss = FocalLoss(ignore_index=self.pad_id, gamma=focal_gamma, reduction='none')
        else:
            self.focal_loss = FocalLoss(ignore_index=self.pad_id, gamma=focal_gamma, reduction='mean')

        self.masked_text = masked_text
        self.masked_text_ratio = masked_text_ratio
        # self.text_mask_token = nn.Parameter(torch.randn(embed_dim))
        self.mask_token_id = len(self.tokenizer.vocab)

        # self.text_position_embed = nn.Parameter(torch.zeros(1, text_length, embed_dim), requires_grad=False)
        self.text_length = text_length

        self.latent_projector_layer = projector_layer
        if self.latent_projector_layer != 0:
            self.latent_projector = [
                nn.Linear(embed_dim, embed_dim),
                nn.ReLU()
            ] * (self.latent_projector_layer - 1)
            self.latent_projector.append(nn.Linear(embed_dim, embed_dim))

            self.latent_projector = nn.Sequential(*self.latent_projector)


        self.initialize_weights()


    def initialize_weights(self):
        # initialization
        # initialize (and freeze) pos_embed by sin-cos embedding
        pos_embed = get_2d_sincos_pos_embed(self.pos_embed.shape[-1], int(self.patch_embed.num_patches**.5), cls_token=True)
        self.pos_embed.data.copy_(torch.from_numpy(pos_embed).float().unsqueeze(0))

        decoder_pos_embed = get_2d_sincos_pos_embed(self.decoder_pos_embed.shape[-1], int(self.patch_embed.num_patches**.5), cls_token=True)
        self.decoder_pos_embed.data.copy_(torch.from_numpy(decoder_pos_embed).float().unsqueeze(0))

        # text_pos_embed = get_1d_sincos_pos_embed_from_grid(self.embed_dim, )
        # torch.nn.init.xavier_normal_(self.text_position_embed) # learnable text position embedding

        # initialize patch_embed like nn.Linear (instead of nn.Conv2d)
        w = self.patch_embed.proj.weight.data
        torch.nn.init.xavier_uniform_(w.view([w.shape[0], -1]))

        # timm's trunc_normal_(std=.02) is effectively normal_(std=0.02) as cutoff is too big (2.)
        torch.nn.init.normal_(self.cls_token, std=.02)
        torch.nn.init.normal_(self.mask_token, std=.02)
        # torch.nn.init.normal_(self.text_mask_token, std=.02)

        # initialize nn.Linear and nn.LayerNorm
        self.apply(self._init_weights)

    def _init_weights(self, m):
        if isinstance(m, nn.Linear):
            # we use xavier_uniform following official JAX ViT:
            torch.nn.init.xavier_uniform_(m.weight)
            if isinstance(m, nn.Linear) and m.bias is not None:
                nn.init.constant_(m.bias, 0)
        elif isinstance(m, nn.LayerNorm):
            nn.init.constant_(m.bias, 0)
            nn.init.constant_(m.weight, 1.0)

    def patchify(self, imgs):
        """
        imgs: (N, 3, H, W)
        x: (N, L, patch_size**2 *3)
        """
        p = self.patch_embed.patch_size[0]
        assert imgs.shape[2] == imgs.shape[3] and imgs.shape[2] % p == 0

        h = w = imgs.shape[2] // p
        x = imgs.reshape(shape=(imgs.shape[0], 3, h, p, w, p))
        x = torch.einsum('nchpwq->nhwpqc', x)
        x = x.reshape(shape=(imgs.shape[0], h * w, p**2 * 3))
        return x

    def unpatchify(self, x):
        """
        x: (N, L, patch_size**2 *3)
        imgs: (N, 3, H, W)
        """
        p = self.patch_embed.patch_size[0]
        h = w = int(x.shape[1]**.5)
        assert h * w == x.shape[1]
        
        x = x.reshape(shape=(x.shape[0], h, w, p, p, 3))
        x = torch.einsum('nhwpqc->nchpwq', x)
        imgs = x.reshape(shape=(x.shape[0], 3, h * p, h * p))
        return imgs

    def random_masking(self, x, mask_ratio):
        """
        Perform per-sample random masking by per-sample shuffling.
        Per-sample shuffling is done by argsort random noise.
        x: [N, L, D], sequence
        """
        N, L, D = x.shape  # batch, length, dim
        len_keep = int(L * (1 - mask_ratio))
        
        noise = torch.rand(N, L, device=x.device)  # noise in [0, 1]
        
        # sort noise for each sample
        ids_shuffle = torch.argsort(noise, dim=1)  # ascend: small is keep, large is remove
        ids_restore = torch.argsort(ids_shuffle, dim=1)

        # keep the first subset
        ids_keep = ids_shuffle[:, :len_keep]
        x_masked = torch.gather(x, dim=1, index=ids_keep.unsqueeze(-1).repeat(1, 1, D))

        # generate the binary mask: 0 is keep, 1 is remove
        mask = torch.ones([N, L], device=x.device)
        mask[:, :len_keep] = 0
        # unshuffle to get the binary mask
        mask = torch.gather(mask, dim=1, index=ids_restore)

        return x_masked, mask, ids_restore, ids_keep

    def forward_encoder(self, x, mask_ratio):
        # embed patches
        x = self.patch_embed(x)

        # add pos embed w/o cls token
        x = x + self.pos_embed[:, 1:, :]

        # masking: length -> length * mask_ratio
        x, mask, ids_restore, ids_keep = self.random_masking(x, mask_ratio)

        # append cls token
        cls_token = self.cls_token + self.pos_embed[:, :1, :]
        cls_tokens = cls_token.expand(x.shape[0], -1, -1)
        x = torch.cat((cls_tokens, x), dim=1)

        if self.extract_multi_level:
            multi_level_feats = []
            # apply Transformer blocks
            for blk_idx, blk in enumerate(self.blocks):
                x = blk(x)
                if blk_idx in [2, 5, 8]:
                    multi_level_feats.append(self.projectors[[2,5,8].index(blk_idx)](x))
            x = self.norm(x)
            multi_level_feats.append(x)

            return multi_level_feats, mask, ids_restore


        # apply Transformer blocks
        for blk_idx, blk in enumerate(self.blocks):
            x = blk(x)
        x = self.norm(x)
        
        return x, mask, ids_restore, ids_keep

    def forward_decoder(self, x, ids_restore):
        # embed tokens
        x = self.decoder_embed(x)
        # non_mask_token = x

        # append mask tokens to sequence
        mask_tokens = self.mask_token.repeat(x.shape[0], ids_restore.shape[1] + 1 - x.shape[1], 1)
        x_ = torch.cat([x[:, 1:, :], mask_tokens], dim=1)  # no cls token
        x_ = torch.gather(x_, dim=1, index=ids_restore.unsqueeze(-1).repeat(1, 1, x.shape[2]))  # unshuffle
        x = torch.cat([x[:, :1, :], x_], dim=1)  # append cls token

        # add pos embed
        x = x + self.decoder_pos_embed

        # apply Transformer blocks
        decoder_feat = []
        for idx, blk in enumerate(self.decoder_blocks):
            x = blk(x)
            if idx == self.mae_decoder_depth // 2:
                decoder_feat.append(x)

        x = self.decoder_norm(x)

        # use the output from decoder to do captioning

        # predictor projection
        x = self.decoder_pred(x)

        # remove cls token
        x = x[:, 1:, :]

        return x, decoder_feat

    def forward_loss(self, imgs, pred, mask):
        """
        imgs: [N, 3, H, W]
        pred: [N, L, p*p*3]
        mask: [N, L], 0 is keep, 1 is remove, 
        """
        target = self.patchify(imgs)
        if self.norm_pix_loss:
            mean = target.mean(dim=-1, keepdim=True)
            var = target.var(dim=-1, keepdim=True)
            target = (target - mean) / (var + 1.e-6)**.5

        loss = (pred - target) ** 2
        loss = loss.mean(dim=-1)  # [N, L], mean loss per patch

        loss = (loss * mask).sum() / mask.sum()  # mean loss on removed patches
        return loss

    def embed_text(self, text):
        batch, device = text.shape[0], text.device

        seq = text.shape[1]

        text_tokens = self.token_emb(text)

        # append text cls tokens
        text_cls_tokens = repeat(self.text_cls_token, 'd -> b 1 d', b=batch)
        text_tokens = torch.cat((text_tokens, text_cls_tokens), dim=-2)

        # create specific mask for text cls token at the end
        # to prevent it from attending to padding
        cls_mask = rearrange(text != self.pad_id, 'b j -> b 1 j')
        attn_mask = F.pad(cls_mask, (0, 1, seq, 0), value=True)

        # go through unimodal layers
        for attn_ff in self.unimodal_layers:
            text_tokens = attn_ff(text_tokens, attn_mask=attn_mask)

        if self.need_uni_2_mul_proj:
            text_tokens = self.uni_2_mul_proj(text_tokens)

        # get text cls token
        text_tokens, text_cls_tokens = text_tokens[:, :-1], text_tokens[:, -1]
        return text_tokens

        
    
    def forward(self, imgs, caption_ids=None, attention_mask=None, mask_ratio=0.75, 
                    freeze_bert=False, teacher_forcing=False, caption_only=False,
                    encoder_only=False, word_weights=None, syn_count=None):
        latent, mask, ids_restore, ids_keep = self.forward_encoder(imgs, mask_ratio)

        if not caption_only:
            pred, decoder_feat = self.forward_decoder(latent, ids_restore)  # [N, L, p*p*3]
            mae_loss = self.forward_loss(imgs, pred, mask)
        else:
            mae_loss = 0.

        if self.latent_projector_layer != 0:
            latent = self.latent_projector(latent)

        # latent: visual info: N, L, C
        # caption_ids: N, Len
        text, labels = caption_ids[:, :-1], caption_ids[:, 1:]

        seq = text.shape[1]
        text_tokens = self.embed_text(text) # N, Len, C

        # create specific mask for text cls token at the end
        # to prevent it from attending to padding
        cls_mask = rearrange(text != self.pad_id, 'b j -> b 1 j')
        attn_mask = F.pad(cls_mask, (0, 1, seq, 0), value=True)
        unimodal_text_tokens = text_tokens
        if not self.less_u:
            for attn_ff, cross_attn in self.multimodal_layers:
                text_tokens = attn_ff(text_tokens, attn_mask=attn_mask[:, :-1, :-1])
                text_tokens = cross_attn(text_tokens, latent)
        else:
            # dim, num_head, 
            for cross_attn1, cross_attn2 in self.multimodal_layers:
                text_tokens = cross_attn1(text_tokens, latent)
                text_tokens = cross_attn2(text_tokens, latent)

        logits = self.to_logits(text_tokens) # N, Len, NVocab
        logits = logits.reshape(-1, len(self.tokenizer.vocab))
        labels = labels.reshape(-1)

        caption_loss = F.cross_entropy(logits, labels, ignore_index=self.pad_id,)


        return mae_loss, caption_loss, None



def mae_vit_small_patch16_dec512d8b(**kwargs):
    model = MaskedAutoencoderViT(
        patch_size=16, embed_dim=384, depth=12, num_heads=6,
        decoder_embed_dim=512, decoder_depth=8, decoder_num_heads=16,
        mlp_ratio=4, norm_layer=partial(nn.LayerNorm, eps=1e-6), **kwargs)
    return model



def mae_vit_base_patch16_dec512d8b(**kwargs):
    model = MaskedAutoencoderViT(
        patch_size=16, embed_dim=768, depth=12, num_heads=12,
        decoder_embed_dim=512, decoder_depth=8, decoder_num_heads=16,
        mlp_ratio=4, norm_layer=partial(nn.LayerNorm, eps=1e-6), **kwargs)
    return model

def mae_vit_large_patch16_dec512d8b(**kwargs):
    model = MaskedAutoencoderViT(
        patch_size=16, embed_dim=1024, depth=24, num_heads=16,
        decoder_embed_dim=512, decoder_depth=8, decoder_num_heads=16,
        mlp_ratio=4, norm_layer=partial(nn.LayerNorm, eps=1e-6), **kwargs)
    return model


def mae_vit_huge_patch14_dec512d8b(**kwargs):
    model = MaskedAutoencoderViT(
        patch_size=14, embed_dim=1280, depth=32, num_heads=16,
        decoder_embed_dim=512, decoder_depth=8, decoder_num_heads=16,
        mlp_ratio=4, norm_layer=partial(nn.LayerNorm, eps=1e-6), **kwargs)
    return model


# set recommended archs
mae_vit_small_patch16 = mae_vit_small_patch16_dec512d8b
mae_vit_base_patch16 = mae_vit_base_patch16_dec512d8b  # decoder: 512 dim, 8 blocks
mae_vit_large_patch16 = mae_vit_large_patch16_dec512d8b  # decoder: 512 dim, 8 blocks
mae_vit_huge_patch14 = mae_vit_huge_patch14_dec512d8b  # decoder: 512 dim, 8 blocks